Question about Feature::generateInitialGuess
TWCV2020 opened this issue · 2 comments
To whom it may concern, I have a question about lines 235 and 236 in Feature::generateInitialGuess in msckf_vio/feature.hpp
It seems that, instead of
A(0) = m(0) - z2(0)*m(2);
A(1) = m(1) - z2(1)*m(2);
we should have
A(0) = m(0) - ( T_c1_c2.translation()(0) / T_c1_c2.translation()(2) ) * m(2);
A(1) = m(1) - ( T_c1_c2.translation()(1) / T_c1_c2.translation()(2) ) * m(2);
The above is obtained by linearizing z2 with respect to depth using first-order Taylor expansion around depth = 0. Please let me know if my understanding is correct. Much appreciated!
I am not sure if linearization is necessary in this case.
If I remember correctly, the initial guess is generated by solving D in Z * z2 = R * D * z1 + t (D and Z are the depth of the feature in the two frames).
The last row of the equation gives Z = D * m(2) + t(2) where m = R * z1.
Using Z in the first two rows should agree with the code.
Got it. I was directly linearizing the right hand side of
z2(0) = (D * m(0) + t(0))/(D * m(2) + t(2))
z2(1) = (D * m(1) + t(1))/(D * m(2) + t(2)),
but we can multiply both sides with (D * m(2) + t(2)) as you explained.
Thank you for the clarification!